Do I have to download all three checkpoints?

No. Only the checkpoint you request is downloaded: the English model at the root is about 808 MB, and the multilingual or typed-decisions builds live in subfolders at roughly 647 MB. Router keeps two checkpoints resident by default so a language switch normally costs only script detection.

Does it expose the same API as Jev?

The optional serve extras do. Install laya[serve] and run laya-serve, which exposes POST /v1/systemone with the same request and response shape, so existing TypeSafe clients work after a base-URL change. Without LAYA_API_KEY set it has no authentication, so do not expose it publicly.

How accurate are the base weights out of the box?

On the typed-decisions benchmark the base English checkpoint scores 0.362 against a 0.461 majority-class floor — the card calls it a fast base to specialise, not a zero-shot decision engine. The 0.766 figure belongs to the fine-tuned laya-typed-decisions checkpoint.

How do I install it, and what are the optional extras?

pip install laya on Python 3.10 or newer. Extras add the HTTP server, an MCP server, LangChain and LangGraph, ONNX Runtime and a TileLang GPU fast path. For a single pipeline, laya.load() on the repo root or a subfolder runs the checkpoint directly; if that call hangs, run with USE_TF=0 because transformers probes for TensorFlow at import.

How long a document can it read?

The multilingual build ships with a 1,024-token limit that cuts long documents off, so pass max_len=8192. The card's table shows 16 to 18 of 20 requests correct with up to about 4,000 tokens of text and 8 to 17 of 20 beyond that, with a 4,000-token input taking about 1.7 s on an Apple GPU. Name the checkpoint explicitly for long mostly-English text or the router sends it to English.

Where does Laya lose to Jev, and where does it win?

Routed Laya leads on typed-decisions at 0.766 against 0.727, AG News 0.950 against 0.910 and DAIR Emotion 0.595 against 0.480, with ECE 0.081 against 0.246 and p50 latency 32.8 ms against 236-276 ms. Jev leads on high-cardinality choice: Banking77 0.870 against 0.425, because 77 options share a fixed token budget. Jev figures are third-party published.

What failure modes does the card warn about?

The English checkpoint collapses on non-Latin scripts — Khmer scores 0.000 accuracy at 0.952 confidence, so confidence gating cannot save you. A noul question can follow its false/true labels instead of the state and return a confident wrong answer, action.act_probability reads 1.0 for almost everything and should be ignored, and ordinal score is the weakest primitive at 0.372 on SST-5.

CommunityJev Models & Reimplementations

Laya on Hugging Face

The Hugging Face hub for Laya's open System 1 checkpoints: the 421M English model at the repo root, a 322M multilingual build in a subfolder, and the typed-decisions fine-tune, all Apache-2.0.

See highlights

Open huggingface.co

Published by: Community
Author: Convai Innovations
Last verified: Oct 2026

Tagscommunity open-models local classification

Works forJev for Classification Jev for Routing Jev for Moderation

Used inJev for Software & DevTools

Languages: Python
Capabilities: Open sourceSelf-hostedRuns locally

Highlights

One repo carries all three checkpoints: the 421M English model at the root and multilingual and typed-decisions builds in subfolders.
Router mode picks a checkpoint by Unicode script before the forward pass, so non-Latin text never reaches the English encoder.
The bundled laya-serve exposes the same POST /v1/systemone shape, so existing clients switch by base URL alone.
Apache-2.0 weights with 41,468 downloads last month and a 5.44k-like community at the time of writing.

Quickstart

python

from laya import Router

router = Router(preload=True)
result = router.predict(
    "We were billed twice for March. Please refund the duplicate.",
    {
        "department": {
            "type": "choice",
            "instructions": "Which department should handle this?",
            "criteria": {"billing": "invoices, refunds", "technical": "bugs, outages"},
        },
        "churn_risk": {"type": "noul", "instructions": "Does the user threaten to cancel?"},
    },
)
print(result["answers"]["department"]["choice"])

Watch out

Base checkpoints sit near chance zero-shot on typed-decisions (0.362 against a 0.461 majority-class floor); the 0.766 belongs to the fine-tuned checkpoint. The models run over-confident until you fit temperatures, and the root checkpoint is English-only unless you route to the multilingual build.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

LinkedIn: What multimodal decision models actually change, and what they do not.

LinkedInJevDirectory.org

Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.

jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker News

Hacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.

1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker News

Hacker News: A one-token logprob wrapper that also classifies webcam frames.

allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker News

Related terms

Glossary definitions related to this entry.

More Jev resources like this

Julia 1

Supersonic Labs' 144.3M-parameter decision model on mmBERT-small: choice, score, and noul answers from supplied options in one pass, 73.15% on 2,000 typed decisions, Apache-2.0.